ResearchBenchmarksMathematical Reasoning on In-domain math benchmarks OverallFollow62.1AccuracyGRPO53.98856.09458.260.306Apr 25, 2026Evaluation ResultsMethodMethodLinksAccuracyPrediction Calibration ErrorExpected Calibration ErrorGRPOConf.=probeConf.=probe2026.0462.10.0650.13C^2GSPGConf.=logitsConf.=logits2026.0461.80.2110.232RLCMConf.=probeConf.=probe2026.0461.80.0360.091RLCRConf.=verbalConf.=verbal2026.04610.1890.189BaseConf.=probeConf.=probe2026.0454.30.0660.101